Quantifier elimination, amalgamation, deductive interpolation and - PowerPoint PPT Presentation

Quantifier elimination, amalgamation, deductive interpolation and Craig interpolation in many-valued logic Franco Montagna, first part in collaboration with Tommaso Cortonesi and Enrico Marchioni

Definition . A logic L has the deductive interpolation property (DIP) if for any set Σ of formulas and for every formula φ , if Σ ⊢ L φ , then there is a formula γ (called a deductive interpolant of Σ and φ ), such that Σ ⊢ L γ , γ ⊢ L φ and all variables in γ are common to Σ and to φ . A logic L has the Craig interpolation property (CIP) if for all formulas ψ, φ , if L ⊢ φ → ψ , then there is a formula γ , (called a Craig interpolant of φ and ψ ) such that L ⊢ φ → γ , L ⊢ L γ → ψ and all variables in γ are common to ψ and to φ .

Example . In classical logic, we have ⊢ ( p ∧ q ) → ( q ∨ r ). A Craig interpolant of p ∧ q and q ∨ r is given by q . q is also a deductive interpolant. For logics with the deduction theorem, CIP and DIP are equivalent. In fuzzy logic, CIP implies DIP but the converse does not hold: � Lukasiewicz logic and product logic have DIP, but not CIP. If φ = p ∧ ( p → q ) and ψ = r ∨ ( r → q ), then φ → ψ is provable in any fuzzy logic, but it does not have a Craig interpolant in � Lulasiewicz or in product logic. Among the most important fuzzy logics, only G¨ odel logic is known to have CIP.

Definition . A V-formation in a class K of algebras of the same type is a system ( A , B , C , i, j ) where A , B , C ∈ K and i, j are embeddings of A into B and into C , respectively. B C i տ ր j A An amalgam in K of a V-formation ( A , B , C , i, j ) is a triplet ( D , h, k ) where D ∈ K and h, k are embeddings of B and C , respectively, into D such that the diagram D h ր տ k B C i տ ր j A commutes. A class K has the amalgamation property (AP) if any V-formation in K has an amalgam in K .

AP implies DIP (in fact, it implies a stronger property, namely, Robinson’s property). In turn, AP is implied by quantifier elimination. A first-order theory T has quantifier elimination (QE) if every formula φ ( x 1 , . . . , x n ) is provably equivalent in T to a quantifier free formula ψ ( x 1 , . . . , x n ). We now present two theorems, the first one is well-known, the second one is a result by Metcalfe, Tsinakis and myself, to (dis)appear. Theorem 1 . If T has QE, then the class of all models of T ∀ , the universal fragment of T , has AP. Theorem 2 . Let V be a variety of representable commutative residuated lattices. If V lin , the class of all chains in V , has AP, then V has AP. From the theorems above we derive the following:

Theorem 3 . Let V a variety of commutative and representable residuated lattices. Suppose that some subclass K of V lin , enjoys the following properties: (1) K is elementary (first-order axiomatizable). (2) Th ( K ) has QE. (3) Every algebra in V lin can be extended to an algebra in K . Then V has AP.

Didactical examples . (1) The class of all divisible abelian o-groups has QE. Every abelian o-group embeds into a divisible abelian o-group. Hence: The class of commutative ℓ -groups has AP. (2) Every MV-chain embeds into a divisible MV-chain and divisible MV- chains have QE. Hence, the class of MV-algebras has AP, see [MuAP] and � Lukasiewicz logic has DIP. A similar result holds for the class of product algebras. (3) The class of densely ordered G¨ odel chains has QE. Every G¨ odel chain embeds into a densely ordered G¨ odel chain. Hence: The class of G¨ odel algebras has AP and G¨ odel logic has DIP. Since G¨ odel logic has the deduction theorem, it also has CIP.

What about BL-algebras? By Theorem 2, we can restrict ourselves to BL-chains. Wanted: a class of BL-chains K such that: (1) Th ( K ) has QE. (2) Every BL-chain embeds into an algebra from K .

In [CMM], we found two examples of such classes, namely: (1) The class of strongly dense BL-chains, that is, the class of BL-chains which are ordinal sums of divisible MV-algebras and the order of components is dense with minimum and without maximum. (2) The class of BL-chains which are ordinal sums of divisible MV-algebras and the order of components is discrete with minimum and without maximum. In this second case, in order to have QE we need to add two new primitives: the function s associating to every element a < 1 the minimum idempotent strictly greater than a , and the function p associating to every a not in the first component the minimum of the component immediately below the one a belongs to.

The class K of strongly dense BL-algebras has QE and every BL-chain A embeds into a chain in K . Indeed, embed the order I of components of A into a dense order J . Then replace every component A i of A by a divisible MV-algebra B i in which A i embeds, and for every j ∈ J \ I add a divisible MV-chain as a new component. Therefore: Theorem 4 . The class of BL-algebras has AP, and BL has DIP.

Craig interpolation . We have seen that none of � Lukasiewicz logic, product logic or BL has CIP. Apart from G¨ odel logic (and classical logic) the most interesting fuzzy logics do not have CIP. For instance, Nilpotent minimum NM, the logic induced by the t-norm x ∗ y = 0 if x + y ≤ 1 and x ∗ y = min { x, y } otherwise, has AP, but not CIP. In [BV], it is shown that both divisible � Lukasiewicz logic � L div and product logic with n th roots Π root have CIP. To prove this, they use an extension of � L div (resp., of Π root ) with propositional quantifiers, and they show that such extensions have QE. Thus a Craig interpolant of φ ( P, Q ) and ψ ( Q, R ), where P, Q, R disjoint sequences of variables, is obtained by eliminating quantifiers in either ∃ P ( φ ( P, Q ), or in ∀ R ( ψ ( Q, R ). Such interpolants are called uniform (the first one only depends on φ and the second one only depends on ψ ).

� L div and Π root are conservative extensions of � Land of Π, respectively. Hence, the following problem arises: Problem . Given a fuzzy logic L which does not satisfy CIP, find a conservative extension of it which satisfies CIP. The argument used by Baaz and Veith shows that for our problem it suffices to find a conservative extension L’ of L with such that: (a) The extension QL’ of L’ by propositional quantifiers has QE. (b) QL’ is a conservative extension of L. (Warning: it is possible that L’ is conservative over L, but QL’ is not conservative over L’).

Our method only works for ∆-core fuzzy logics, roughly, for logics having the Baaz-Monteiro operator ∆. Then under suitable additional assumptions, it is possible to interpret QL’ into the first order theory, Th ( L ′ ), of all L’-chains and viceversa in such a way that Th ( L ′ ) has QE iff QL’ has QE. In this way, we arrive to the following general theorem:

Theorem 5 . Let L’ be a conservative extension with ∆ of a fuzzy logic L, and let L ′ be the class of L’-chains. Suppose that: (a) Th ( L ′ ) is axiomatizable by universal formulas and has quantifier elimina- tion. (b) L ′ has a model A which is complete with respect to the order and a prime model B . Then: (1) QL’ has QE. (2) QL’ is conservative over L. (3) L’ has CIP.

Applications . Any finitely valued fuzzy logic falls under the scope of The- orem 5 . If we add a constant for each truth value and the Baaz-Monteiro operator ∆, we obtain a conservative extension with CIP. But this example is straightforward, and ∆ is not necessary in this case. A more interesting example is the following: let BL’ be the logic of BL- chains which are ordinal sums of divisible MV-algebras with a discrete order of components added with operators s and p ( s ( x ) is the minimum idempotent strictly above x and p ( x ) is the minimum of the component immediately below the component x belongs to). Then BL’ satisfies the conditions of Theorem 5. Hence, BL’ has CIP and is conservative over BL. A third example is the following: NM, the Nilpotent Minimum logic, is strongly complete wrt the class of all discretely ordered NM-chains. Now let us add to NM a symbol 1 2 for the fixpoint of the negation, and two operators s and p such that, if x < 1, then s ( x ) is the minimum element > x and if 0 < x < 1, then p ( x ) is the greatest element strictly below x ). In this way we obtain a logic L’ which satisfies all assumptions of Theorem 5. Hence, L’ is a conservative extension of NM which satisfies CIP.

References . [AM] P. Aglian´ o, F. Montagna, Varieties of BL-algebras I: general properties , Journal of Pure and Applied Algebra 181 (2003), 105-129. [BHMV] M. Baaz, P. H´ ajek, F. Montagna, H. Veith, Complexity of t-tautologies , Annals of Pure and Applied Logic 113 (2002), 3-11. [BKW] A. Bigard, K. Keimel and S. Wolfenstein, Groupes at anneaux retic- es , Lecture Notes in Mathematics, 608 , Springer Verlag, Berlin, 1977. ul´ [BM] Manuela Busaniche, Daniele Mundici: Geometry of Robinson consisten- cy in Lukasiewicz logic , Ann. Pure Appl. Logic 147(1-2): 1-22 (2007). [Ca] X. Caicedo. Implicit operations in MV-algebras and the connectives of Lukasiewicz logic , In Algebraic and Proof-Theoretic Aspects of Non-classical � Logics, 50–68, 2006. [CEGT] R. Cignoli, F. Esteva, L. Godo, A. Torrens, Basic fuzzy logic is the logic of continuous t-norms and their residua , Soft Computing 4 (2000), 106-112. [CDM] R. Cignoli, I.M.L. D’Ottaviano, D. Mundici, Algebraic foundations of many-valued reasoning , Kluwer, 2000.